[zhangzhenghao]

Summary

Follow llama.cpp's architecture-selected test runner and per-model tokenizer registration pattern: one shared runner, with architecture fixtures and expectations registered separately.

• Introduce tests/models/model_tests.py with --arch, --suite, --case and fixture-free --list.
• Centralize subprocess execution, exit/diagnostic assertions, JSON vector/frame comparisons, WAV generation and synthetic GGUF metadata rewriting in model_contract.py.
• Replace six standalone CLAP/CLAP2/CLaMP3/MERT/MERT2 scripts with architecture adapters; this is not a dispatcher to the old scripts.
• Register existing CTest cases through one helper, preserving names, option gates, dependencies, fixture setup and timeouts.
• Retain every migrated mutation and numerical threshold, including positive-only MERT rejection exits, strict CLAP error bounds, inclusive MERT2 bounds and its 120-second invocation deadlines.
• Keep specialized official numerical oracles, native C API tests, FFT/RoPE/tokenizer internals, bundled pipelines and MuQ-specific contracts separate and unchanged.
• Add 21 framework regressions and document adding/selecting architecture contracts.

Reference: https://github.com/ggml-org/llama.cpp/blob/master/tests/test-llama-archs.cpp and https://github.com/ggml-org/llama.cpp/blob/master/tests/CMakeLists.txt.

Verification

• Default static Release build: 36/36 CTests passed.
• Default shared Release build: 35/35 CTests passed (allocation-interposition test remains static-only).
• Optional CLAP/CLaMP3/MERT2 parity, native, tokenizer and migrated CLI/metadata contracts with freshly generated fixtures: 10/10 CTests passed.
• Framework regressions: 21 passed.
• Deliberately corrupted a MERT2 pooled reference: real CLI succeeded, unified test correctly exited 1 with architecture/suite/case, numerical error and captured command output.
• Independent semantic review: no blocking findings; MERT base config and all 31 native mutations mechanically compared with the baseline.
• Python compilation and git diff --check passed.

An initial run against historical CLaMP3 fixtures failed because those GGUFs lacked bundled MERT metadata; the baseline script fails identically. Regenerating fixtures with the current parity pipeline passed all optional checks.

GitHub Actions on b1c62939daf9b4f05850553e1c6753cc0754f1e5: all 9 checks passed — clang-format; Linux x64/ARM64, macOS ARM64 and Windows x64/ARM64 builds/tests; Rust checks on Linux/macOS/Windows. Verified CI run: https://github.com/gorse-io/clap.cpp/actions/runs/42498659326.